Briefing · 24-hour window · 29 sources tracked

What moved in AI.

A quiet Sunday. Ethan Mollick's SatBench shows labs are running out of benchmarks: 63% of the ones they cite are now effectively beaten. Claude's auto mode permission system has reached everyday chat, and fresh UK data shows Claude's share of AI spending quadrupling even as agent shopping stays tiny.

By the numbers

The day in three figures.

63%

of the 173 benchmarks AI labs cite most are now saturated (best score at 85% or above).

Source: SatBench, Ethan Mollick

1 in 20

UK adults now pay for AI, up from 1 in 100 two years ago, spending over $1.3 billion a year.

Source: Exponential View #610

4

Uber rides booked by AI agents out of 320 million during Halloween week.

Source: Uber, via Exponential View

The news

Five things to know.

01 · Research / BenchmarksRelevant to you

SatBench: 63% of the AI benchmarks labs cite are now beaten

What happened
At 00:46 Paris time, Wharton professor Ethan Mollick published SatBench, which tracks how fast major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the labs since 2023, 63% are now saturated, and saturation is happening faster each year. Mollick says Fable 5.2 built the analysis in about four prompts.
Why it matters
Public leaderboards are losing their power to separate frontier models, so model choices increasingly rest on your own task evaluations. Relevant to you: benchmark claims in vendor pitches deserve less weight than a test on your own processes.
02 · Claude / AgentsRelevant to you

Claude's auto mode now runs in everyday Claude chat

What happened
Simon Willison reported on Saturday evening that regular Claude chat now uses auto mode, in which a background classifier, rather than a prompt to the user, approves or blocks actions, and that it takes the user's own messages into account. Anthropic had said on October 22 that auto mode would become the default for new sessions in most Claude apps on Pro, Max and Team plans.
Why it matters
Fewer approval prompts make Claude faster on routine work, but it can now occasionally refuse an action on its own. Relevant to you: worth checking what Claude may do unprompted before wider rollout in a team.
03 · Adoption / AgentsRelevant to you

Claude's share of UK AI spending quadruples while agent shopping stays tiny

What happened
Exponential View #610 (Azeem Azhar, this morning) reports that one in twenty UK adults now pays for AI, spending more than $1.3 billion a year, and that Claude's share has quadrupled in a year. It also cites Uber: only 4 of 320 million Halloween-week rides were booked by AI agents. Partly paywalled.
Why it matters
Agents pay off on tedious, repetitive tasks rather than simple ones where an app is faster, a useful filter for choosing enterprise use cases. Relevant to you: the adoption numbers support Claude's business case.
04 · Creative AI / Models

Fable 5.2 makes a 4K Blender film from a single prompt

What happened
Riley Brown shared The Fable of Clawd, a 2-minute 4K film made entirely in Blender, the open-source 3D software, from one prompt to Claude Fable 5.2. Claude built every character, prop and scene by writing Python, in about 4.5 hours.
Why it matters
It shows a frontier model running a long, multi-step production job end to end through code, the same pattern that lets agents drive enterprise tools.
05 · Open models / Tools

Datasette agents can now run on local models

What happened
Simon Willison released datasette-llm 0.1a11, which adds support for any OpenAI-compatible endpoint, so Datasette's agent can run on a local model through LM Studio, Ollama or llama.cpp. llama.cpp itself reached version 1.0.0 on November 13 with a formal API stability promise.
Why it matters
Local models are becoming a drop-in option for data tools, useful where data cannot leave the company's machines.

In brief

Also worth a look.

Unread today: X was read signed out, so some posts may be missing. Not read directly: One Useful Thing, How to AI, Behind the Craft, Forward Future, darioamodei.com, blog.samaltman.com, karpathy.bearblog.dev, The Neuron site, RoboNuggets free community.